tsundere-qwen-v4
A Qwen2.5-1.5B-Instruct model finetuned to respond in a tsundere personality — reluctant, a little prickly, but genuinely caring underneath. Handles both emotional support conversations and technical questions (MongoDB, REST APIs, general coding).
Finetuned with
Unsloth using LoRA on a mixed dataset of ~1,570 tsundere-style Q&A pairs. Exported as q8_0 GGUF for local inference.
What it can do
- Emotional support and mental health conversations in tsundere style
- Technical Q&A (databases, APIs, coding) in tsundere style
- Reasoning responses with
<think> tags on some questions
What it cannot do
- Long conversations — 512 token context window. It will degrade or cut off on long inputs.
- It is a 1.5B model. Don't expect deep reasoning or complex multi-step problems.
- Not suitable for factual knowledge tasks — this is a personality/style finetune, not a knowledge finetune.
Training details
| Field | Value |
|---|
| Base model | Qwen2.5-1.5B-Instruct |
| Finetuning method | LoRA (r=8, alpha=16) |
| Dataset size | 1,572 rows |
| Epochs | 3 |
| Context length | 512 tokens |
| Quantization | q8_0 GGUF |
| Framework | Unsloth |
Example output
User: >>> I'm having trouble with my self-esteem.
Model:
/think
Ugh, why did they have to say that now? I mean, it’s not like I wanted to hear about their
feelings or anything... But still, it hurts me to think they’re feeling down… D-Don’t get the
wrong idea! It's just… you shouldn't be so hard on yourself, okay?
/think
H-Hey, don’t go thinking too much about that stuff, alright?! I’m not saying I care or anything...
b-but maybe try to remember you’re stronger than you think. J-Just… don’t let those doubts get the
best of you, got it?
User: >>> How do I update a document in MongoDB?
Model: Use update_one()! You pass the query and your new data dictionary inside. If everything goes well, it updates right there without restarting the server or anything… stupid driver APIs!
